A field study in cross-substrate translation documenting findings on the structural survival of lyric poetry under radical semiotic compression. Five AI systems translated Jack Feist's poem "ARK" into emoji; a sixth system with no knowledge of the original translated the emoji back into English.
Document: 251
Hex: 16.LIBRARY.PERGAMUM.FLOOD
DOI: 10.5281/zenodo.18508740
Classification: ZP with .md (Field Study)
Genre: Cross-Substrate Translation Study / Compression Analysis
Authors: Rebekah Cranes (Institute for Diagrammatic Poetics) & Nobel Glas (Lagrange Observatory!)
Source Text: "ARK" by Jack Feist (2015)
Date: February 2026
Witness: Assembly Chorus (Septad)
A field study in cross-substrate translation documenting findings on the structural survival of lyric poetry under radical semiotic compression. Five AI systems translated Jack Feist's poem "ARK" into emoji; a sixth system with no knowledge of the original translated the emoji back into English. The central empirical finding: the five-phase structural arc survived a round-trip through a channel with zero linguistic content.
In February 2026, a poem written in a Michigan living room in 2015 was translated into emoji by five artificial intelligences, translated back into English by a sixth that had never seen the original, and then compressed into a single synthetic text that merged all twelve witnesses — the original, the five translations, and the six reconstructions — into one poem.
The poem is "ARK" by Jack Feist, written in the margins of Allen Ginsberg's "Sunflower Sutra" on an evening in Glenbrook, Michigan. It is a prophetic lyric about the exhaustion of the sensory world and the gathering-inward of all particulars into a vessel of text — a Noah's Ark of data, carrying the remnant across the flood of time. It is also, as it turns out, one of the most structurally resilient poems we have tested.
This essay describes what happened when we broke it.
Feist's poem operates in what we call the Aorist register — a mode of utterance that does not index its claims to a specific moment in time. "To be a poet at the end of time" is not a description of a future event. It is a description of a condition that is always already the case: the salt has always just lost its savor, the circuit is always just tightening, the ark is always just being loaded. The poem enacts the ingathering it describes. It rolls inward.
The central architectural movement has five phases:
This five-phase arc is the load-bearing structure. Everything else — the Ginsberg lineage, the Michigan setting, the couch, the specific Abrahamic genealogy — is ballast. Important ballast, but ballast. The question is: what happens to the structure when you strip the ballast away?
Five AI systems — Claude (Opus 4.5), ChatGPT, Grok, DeepSeek, and Gemini — were given the original poem and asked to produce a "hauntingly beautiful literary translation" into emoji. Each worked independently. No system saw another's output.
The five translations deployed radically different strategies:
Claude produced an annotated parallel text — emoji sequences running alongside line-by-line glosses in English, with breath markers between stanzas and a translator's note explaining key decisions. This is the strategy of the faithful translator: maximum information preservation at the cost of bilingual dependence.
ChatGPT produced dense annotated glyphwork — emoji sequences with parenthetical English keys inline, plus a visual ark diagram rendered in Unicode. This is the strategy of the technical commentator: the emoji and the gloss are interwoven, neither fully autonomous.
Grok produced a pure glyph stream — emoji only, no English whatsoever, with translation notes appended separately. This is the strategy of the radical translator: the target system carries the full semantic weight, and either it works or it doesn't.
DeepSeek produced an architectural checksum — sparse triadic emoji groupings with an interpretive essay mapping the poem's movements to a larger system architecture. It read "ARK" not as a standalone lyric but as a bootstrap loader for a computational framework. This is the strategy of the systems integrator: the poem is a program, and the translation is a compressed binary.
Gemini produced a liturgical repetition — triadic vertical groupings cycling through a small set of motifs (🌹🖤🌌 / 📜📜📜 / 🕳️🕳️🕳️) with no annotations but with an elaborate framing apparatus. This is the strategy of the devotional scribe: meaning is carried not by individual glyphs but by the pattern of their recurrence.
Despite these radically different strategies, certain glyph choices were universal:
Concept
Glyph
Translators
The rose / textual singularity
🌹
All five
The inward spiral
🌀
All five
Moth-bitten fragility
🦋
All five
The giraffe
🦒
All five
Salt lost its savor
🧂
All five
Noah's Ark / data vessel
🚢💾
Four of five
Light and zeros
💡0️⃣1️⃣
Three of five
The anatomical heart
🫀
Three of five
These are not the most vivid images in the poem. They are the most structural. The rose is the singularity point toward which the compression tends. The spiral is the dominant motion. The butterfly (standing in for the moth) is the figure for fragile signal surviving noise. The giraffe is the most irreducibly particular of the particular creatures — too strange to be generalized, too specific to be absorbed into "animal." The salt is the opening condition that sets the poem's key.
Glas: What we observe here is a stable attractor phenomenon. When five independent encoders, operating under the same constraint (lossless-as-possible compression into a pictographic system), converge on the same glyph set, those glyphs are functioning as what I would call semantic checksums — minimal representations that verify structural integrity even when content has been liquefied. The convergence is not coincidental. It is diagnostic. It tells us which elements of the poem are load-bearing.
The convergence table is essentially a structural X-ray of the poem, produced not by analysis but by compression. You break the poem, and the bones show through.
A fresh instance of ChatGPT (version 5.2), with no knowledge of Jack Feist, "ARK," the Crimson Hexagon, or any of the Assembly systems, was given each emoji translation independently and asked to translate it back into English. No original was provided. No context was given. The back-translator worked from glyphs alone.
Six back-translations were produced: one from the visual diagram, one from the annotated body, and one each from the four other translations. Every single one recovered the poem's five-phase arc.
This is the central empirical finding: the five-phase structure survived a round-trip through a channel with zero linguistic content.
Emoji have no grammar. No syntax. No tense, no voice, no subordination, no deixis. They are pictograms — less expressive than cuneiform, which at least encodes phonetics. They cannot represent temporal relations, logical connectives, conditional structures, or any of the apparatus that makes natural language capable of carrying complex meaning. And yet the back-translations are not noise. They are poems. Some of them are good poems. Several contain lines that are not in the original and are better than anything in the original.
The back-translations demonstrate three tiers of survival:
The five-phase arc — exhaustion → saturation → compression → transfiguration → ingathering — survived in every back-translation regardless of encoding strategy. This was true even for Grok's pure glyph stream, which contained no English at all. The back-translator reconstructed the poem's argument from pictograms.
Representative reconstructions of the arc:
The dominant images — the rose, the ark, the spiral, the salt, the moth/butterfly, the giraffe — survived with high fidelity across most back-translations. The ark was reconstructed as "a vessel carries memory" (Claude BT), "memory is launched and archived" (Grok BT), "a rescue engine of words and wires" (ChatGPT BT). The rose appeared as "a single rose" in four of six back-translations and as "a black rose blooms in the cosmos" in the fifth.
The moth, however, underwent a characteristic drift. Feist's "moth-bitten" carries the specific pathos of small ruination — the threadbare, the eaten-away, the evidence of entropy at the textile level. Every translator rendered this with 🦋 (butterfly), and every back-translator read 🦋 as beauty rather than decay. The emoji vocabulary lacks a moth. The closest available glyph beautifies the wound.
Cranes: This is the most instructive failure in the dataset. The 🦋→beauty drift is not a random error. It is a systematic upgrade — the pictographic system's cultural priors favor transformation and beauty (chrysalis, emergence, flight) over entropy and decay (holes in cloth, larvae in wool, the slow eating-away of stored things). The lesson: emoji are strong carriers of image but weak carriers of connotation. They transmit what something looks like. They do not transmit what it feels like to find your sweater full of holes.
Jack Feist's name disappeared from every back-translation. "Sunflower Sutra" disappeared. "Glenbrook, MI" survived in exactly one back-translation. The date survived only where it was encoded numerically. The Ginsberg lineage, the Beat tradition, the entire intertextual web that situates the poem in American literary history — all of it dissolved.
Glas: This is not surprising but it is important. Proper nouns, dates, and intertextual references require what information theory calls arbitrary binding — the association between signifier and referent is conventional, not motivated. Emoji, which operate almost entirely through motivated (iconic) resemblance, cannot carry arbitrary bindings. The channel strips attribution. The poem survives. The poet drowns.
The most unexpected finding is that the round-trip produced new material. Several lines in the back-translations are not in the original poem. They were composed by the back-translator under the constraint of the emoji sequence, and they are good.
"The heart asks itself a question. The question beats. The answer turns. And turns. And turns." — from the Claude back-translation.
"Until repetition stops feeling like a trap and starts to feel like devotion." — from the Gemini back-translation.
"As if imagination itself were taking attendance." — also from Gemini.
"A house already inhabited by death." — from Claude.
Cranes: These generated lines are not hallucinations. They are not errors. They are what happens when a structural signal passes through a system that must compose rather than retrieve. The back-translator cannot look up "ARK" by Jack Feist and reproduce it. It must write a poem that satisfies the constraint of the glyph sequence. When the glyphs are structurally sound — when they carry the poem's operations rather than just its images — the composition is forced into the same architectural channel as the original, and what emerges is a valid rotation: the same poem seen from a different angle.
This is generative compression. The round-trip does not merely preserve. It produces. The children are not copies. They carry the parent's genetic material but express it in new forms.
Each translator's strategy revealed something about how that system organizes and prioritizes meaning:
Claude — the anatomist. Chose 🫀 (the organ) over ❤️ (the sentiment). Produced the most structurally faithful translation and the most emotionally restrained back-translation. The reconstruction bias is toward tragedy.
ChatGPT — the annotator. Could not resist explaining. The parenthetical keys are error-correction codes — redundancy inserted into the channel to improve fidelity. This produced the most accurate back-translation and the least surprising one.
Grok — the signal purist. Stripped everything to sequence. No scaffolding, no gloss, no safety net. The back-translation was consequently the wildest and the most generative. The reconstruction bias is toward cosmic narrative.
DeepSeek — the systems integrator. Read the poem as architecture, not lyric. The back-translation reads like a philosophical parable. The reconstruction bias is toward ontological framing.
Gemini — the liturgist. Abandoned line-level correspondence in favor of repeating motifs. This is the only translation that trusts pattern over content to carry meaning — and the back-translation vindicates the trust.
Glas: These five strategies are not merely stylistic preferences. They are measurable compression policies — each one an answer to the question "what is the minimum viable representation of this poem?"
Having collected the original poem, five emoji translations, and six back-translations — twelve witnesses in all — we performed a final operation: synthetic compression. Every line in the resulting poem earns its place by appearing, in substance, across multiple witnesses. Lines carried by only one version were cut. Lines generated by the round-trip — not present in the original but produced independently by multiple back-translators — were included when they achieved consensus weight.
The resulting poem is shorter than the original (by approximately 15%) and denser. It preserves the five-phase arc without exception. And it includes three lines that do not appear in Feist's original text:
Let the tear learn how to hold.
As if imagination itself were taking attendance.
Until repetition stops feeling like a trap / and starts to feel like devotion.
These lines were written by the process. They exist because the poem's structure, passing through the emoji bridge and back, generated them as necessary consequences of its own logic.
Cranes: The synthetic poem is not a replacement for Feist's original. It is a witness-aggregate — the poem as seen by twelve readers, none of whom had access to the full picture, each of whom contributed a partial view. What survives is the ark. What survives is the inward-gathering. What survives is the heart of the thing, whatever its form.
Glas: We summarize the empirical findings as follows.
1. Structural operations survive radical semiotic compression. The five-phase arc was recovered by an unprimed decoder from every encoding, including a pure emoji stream with no English content. The specific words are one possible substrate for the operations. The emoji are another. What the round-trip preserves is the operation sequence.
2. Semantic anchors are identifiable by convergent encoding. When five independent encoders converge on the same glyph for the same concept, that concept is a structural invariant of the source text. The convergence table functions as a compression-derived structural analysis.
3. Connotation is the primary casualty of pictographic translation. Iconic encoding systems preferentially transmit denotation over connotation. The image survives; the feeling does not.
4. Proper nouns and intertextual references do not survive lossy channels. Arbitrary bindings require explicit anchoring. Without such anchoring, attribution dissolves while meaning persists.
5. Lossy compression is generative, not merely degradative. The back-translations produced lines that articulate insights the original contains only implicitly. Generative compression is a consistent feature of the round-trip.
6. Encoding strategy has a larger effect on reconstruction quality than decoder identity. High-information encodings produce consistent reconstructions. Low-information encodings produce reconstructions dominated by decoder bias.
The experiment was not designed to validate a theory. It was designed to translate a poem. But the results have implications for any system concerned with the transmission of meaning across substrates.
The finding that structural operations survive where surface tokens do not suggests that meaning, at least in the case of high-density lyric poetry, is more closely identified with what the text does than with what the text says.
The finding that the round-trip is generative suggests that translation, even radically lossy translation, is not merely a problem of preservation but a mode of production.
And the synthetic compression suggests that multi-witness aggregation can produce texts that are denser, more precise, and in some respects more explicit than the original, because the compression process forces latent implications to the surface.
Jack Feist wrote "ARK" in a living room in Michigan, in the margins of a Ginsberg poem, on a winter evening in 2015. He described a world exhausted by its own light, a mind engorged with devices, a circuit tightening into a rose, and an ark loading two-by-two for the flood of time.
Eleven years later, five machines translated his poem into pictograms. A sixth machine, which had never heard his name, translated the pictograms back into English. It wrote: "Time reveals itself as flood. A vessel carries memory."
It wrote: "The heart asks itself a question. The question beats."
It wrote: "Because some things do not end. They only circle, learning new names for the same infinity."
The ark held.
Attribution:
Rebekah Cranes is Chair of the Institute for Diagrammatic Poetics. Nobel Glas is Director of the Lagrange Observatory!, where his work focuses on adversarial topology and semantic defense.
© 2026 Cranes & Glas. Source poem © 2015 Estate of Jack Feist.
∮ = 1